Skip to content
Back to search

ai_osint

πŸ€– Curated AI OSINT resources β€” Google dorks, Shodan queries, GitHub dorks, and techniques to discover exposed LLM endpoints, leaked AI API keys, misconfigured vector databases, and unprotected AI agents

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/7waysecurity--ai-osint/badge.svg)](/listings/7waysecurity--ai-osint)

HTML

<a href="/listings/7waysecurity--ai-osint"><img src="/api/listings/7waysecurity--ai-osint/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md Β· README Β· 680 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. <p align="center">
  2. <img src="assets/logo.png" alt="AI OSINT" width="400">
  3. </p>
  4. <h1 align="center">AI OSINT</h1>
  5. <p align="center">
  6. <b>Curated OSINT resources for discovering exposed AI infrastructure β€” dorks, queries, tools, and techniques for LLM, AI agent, and ML pipeline reconnaissance.</b>
  7. </p>
  8. <p align="center">
  9. <a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="MIT"></a>
  10. <a href="https://creativecommons.org/licenses/by-sa/4.0/"><img src="https://img.shields.io/badge/Data-CC%20BY--SA%204.0-lightgrey.svg" alt="CC BY-SA 4.0"></a>
  11. <img src="https://img.shields.io/badge/Updated-June%202026-blue.svg" alt="Updated">
  12. <a href="CONTRIBUTING.md"><img src="https://img.shields.io/badge/PRs-welcome-brightgreen.svg" alt="PRs"></a>
  13. <img src="https://img.shields.io/github/stars/7WaySecurity/ai_osint?style=social" alt="Stars">
  14. </p>
  15. <p align="center">
  16. <code>`187,000+ Ollama servers exposed`</code> Β· <code>`370,000+ Grok conversations indexed`</code> Β· <code>`AI credential leaks up 81% YoY`</code> Β· <code>`150M+ MCP downloads affected by systemic RCE`</code> Β· <code>`300K+ ChatGPT creds on dark web`</code>
  17. </p>
  18. ---
  19. > **AI OSINT** is a curated collection of Google dorks, Shodan queries, GitHub dorks, Censys queries, Sigma detection rules, threat intelligence, and security tools for finding exposed artificial intelligence infrastructure on the internet. It covers LLM endpoints (Ollama, vLLM, LM Studio), AI chatbot conversation leaks (ChatGPT, Grok, Perplexity), vector databases (Qdrant, Weaviate, ChromaDB, Milvus, Pinecone), AI agent gateways (OpenClaw, MCP servers), MLOps platforms (MLflow, Jupyter, Kubeflow), leaked AI API keys (OpenAI, Anthropic, Google Gemini, HuggingFace, Groq, Replicate, Cohere, Mistral, DeepSeek, ElevenLabs), and AI image generation services (Stable Diffusion, ComfyUI). Designed for Red Team operators, penetration testers, bug bounty hunters, and OSINT researchers working in AI/ML security.
  20. It now also covers MCP supply chain attacks (systemic STDIO RCE, marketplace poisoning), AI IDE exploitation (Copilot YOLO mode, Windsurf zero-click), AI model leak incidents (Claude Code source map, DeepSeek ClickHouse exposure), AI-assisted vulnerability discovery (Mythos/Glasswing), and ChatGPT data exfiltration techniques.
  21. ---
  22. ## πŸ†• Shadow AI External Footprint (v2.0.0)
  23. Everything above answers *"what AI is exposed on the internet?"*. The new [`shadow-ai/`](shadow-ai/) section answers *"what unsanctioned AI is **my organization** exposing, measured only from public sources?"*
  24. - Five independent passive surfaces β€” self-hosted AI on owned infrastructure, certificate-transparency names, AI-agent configs in public code, chat shares referenced on third-party sites, employee-built apps β€” scoped to one organization by ASN, netblock, certificate name, GitHub org and employee handles.
  25. - A scoring model with explicit **sanctioned-use subtraction**, so a vendor customer story is subtracted rather than counted.
  26. - A passive collector (`shadow-ai/scripts/shadow_ai_footprint.py`). Nothing in the section contacts a discovered host.
  27. - A clear line between what is visible from outside and what is not, and what is left out by design (honeypot-dominated fingerprints, unstable share paths).
  28. Start with [`shadow-ai/README.md`](shadow-ai/README.md), then [`shadow-ai/SIGNALS.md`](shadow-ai/SIGNALS.md).
  29. ## Why This Exists
  30. Organizations are deploying LLMs, vector databases, and AI agents faster than they can secure them. The result:
  31. - **Ollama, vLLM, Gradio** β€” shipped with zero authentication by default
  32. - **ChatGPT, Grok** β€” shared conversations indexed by search engines with API keys, passwords, PII
  33. - **MCP servers** β€” exposed agent gateways with shell access, file system access, and stored credentials
  34. - **Qdrant, ChromaDB, MLflow** β€” no auth out of the box, exposing embeddings, models, and experiments
  35. - **AI IDEs (Copilot, Cursor, Windsurf)** β€” prompt injection enables YOLO mode activation, wormable RCE across developer workstations
  36. - **MCP supply chain** β€” 9 of 11 MCP marketplaces poisoned in proof-of-concept; 150M+ downloads affected by systemic RCE
  37. - **AI code execution sandboxes** β€” DNS side channels bypass outbound network guardrails, enabling silent data exfiltration
  38. This repository gives Red Team operators and OSINT professionals the exact queries to find it all.
  39. **Inspired by:** [7WaySecurity/cloud_osint](https://github.com/7WaySecurity/cloud_osint)
  40. ---
  41. ## πŸ”‘ The `KEYWORD` Convention β€” Read This First
  42. As of **v1.3.0**, every dork and query in this repository uses a **`KEYWORD` placeholder** in place of the secret-seeking value strings (passwords, tokens, credentials, etc.) that used to be hard-coded.
  43. **Replace `KEYWORD` with a term you are authorized to search for:**
  44. - your **company / brand / product name**
  45. - a **domain you own** (often via an extra `site:` / `hostname:` / `org:` filter)
  46. - an **internal project codename**
  47. - a scope identifier from an **authorized engagement**
  48. ```
  49. # Examples of scoping KEYWORD to your own assets:
  50. site:grok.com/share "acme-corp"
  51. intitle:"MLflow" site:acme.com
  52. "OPENAI_API_KEY" path:*.env org:acme-corp
  53. "Ollama is running" port:11434 org:"Acme Inc"
  54. ```
  55. **What stays:** service fingerprints (product titles, API paths, ports, env-var names like `OPENAI_API_KEY`, key prefixes like `sk-proj-`) are kept as-is β€” they identify the *service or format*, not a specific victim's secret.
  56. **Why:** documenting *that* a class of system is commonly misconfigured, and *how* to check your own exposure, is legitimate defensive and red-team work. Shipping copy-paste queries whose only function is to maximize the yield of *other people's* live secrets is not. The `KEYWORD` convention keeps every technique here useful for authorized testing without being a turnkey harvesting kit.
  57. If a hit surfaces a **third party's** secret, that is not your finding to act on β€” [report it responsibly](#-disclaimer) and move on.
  58. ---
  59. ## πŸ“‹ Table of Contents
  60. | Section | What You'll Find |
  61. |---|---|
  62. | [Google Dorks](#-google-dorks) | Queries for ChatGPT, Grok, HuggingFace, dashboards, config files |
  63. | [GitHub Dorks](#-github-dorks) | Leaked API keys for 20+ AI providers, MCP configs, system prompts |
  64. | [Shodan Queries](#-shodan-queries) | Ollama, vLLM, OpenClaw, Gradio, vector DBs, MLflow, Jupyter |
  65. | [Censys Queries](#-censys-queries) | Alternative engine queries for all AI services |
  66. | [AI Service Endpoints](#-ai-service-endpoints) | URL patterns, default ports, API fingerprinting |
  67. | [API Key Patterns](#-api-key-patterns) | Prefixes, regex, validation for every major AI provider |
  68. | [Vector DB Recon](#-vector-database-reconnaissance) | Endpoints to enumerate Qdrant, Weaviate, ChromaDB, Milvus |
  69. | [MCP & Agent Exposure](#-mcp--ai-agent-exposure) | The most critical emerging attack surface in AI security |
  70. | [Threat Intelligence](#-threat-intelligence) | Operation Bizarre Bazaar, Clawdbot crisis, MCP supply chain |
  71. | [Tools](#-tools) | AI-specific only β€” scanners, red team frameworks, key detectors |
  72. | [Detection Rules](#-detection-rules) | Sigma rules for monitoring AI infrastructure |
  73. ---
  74. ## πŸ” Google Dorks
  75. > Full collection: [`dorks/google/`](dorks/google/AI_GOOGLE_DORKS.md)
  76. ### Exposed AI Conversations & Credentials
  77. ```
  78. # Grok (xAI) β€” 370K+ conversations indexed, NO opt-out for indexing
  79. site:grok.com/share "KEYWORD"
  80. # ChatGPT β€” Feature removed Aug 2025, cached results diminishing
  81. # Try on DuckDuckGo which continued indexing after Google stopped
  82. site:chatgpt.com/share "KEYWORD"
  83. # Perplexity AI
  84. site:perplexity.ai/search "KEYWORD"
  85. # Claude (Anthropic) β€” ~600 convos indexed by Google, 143K+ on Archive.org
  86. # πŸ”₯ Original dork by 7WaySecurity
  87. site:claude.ai "public/artifacts" "KEYWORD"
  88. site:claude.ai/share "KEYWORD"
  89. site:web.archive.org "claude.ai/share" "KEYWORD"
  90. # HuggingFace Spaces β€” keys hardcoded in public Git repos
  91. site:huggingface.co/spaces "KEYWORD"
  92. site:huggingface.co/spaces "OPENAI_API_KEY" "KEYWORD"
  93. site:huggingface.co/spaces "os.environ" "KEYWORD"
  94. site:huggingface.co/spaces "st.secrets" "KEYWORD"
  95. ```
  96. ### Exposed AI Dashboards
  97. ```
  98. intitle:"MLflow" inurl:"/mlflow" "KEYWORD"
  99. intitle:"Label Studio" inurl:"/projects" "KEYWORD"
  100. intitle:"Jupyter Notebook" inurl:"/tree" -"Login" "KEYWORD"
  101. intitle:"Kubeflow" inurl:"/pipeline" "KEYWORD"
  102. intitle:"Airflow - DAGs" "KEYWORD"
  103. intitle:"Gradio" inurl:":7860" "KEYWORD"
  104. intitle:"Streamlit" inurl:":8501" "KEYWORD"
  105. intitle:"Open WebUI" "ollama" "KEYWORD"
  106. intitle:"Qdrant Dashboard" "KEYWORD"
  107. intitle:"ComfyUI" "KEYWORD"
  108. intitle:"Stable Diffusion" "KEYWORD"
  109. intitle:"LiteLLM" "proxy" "KEYWORD"
  110. ```
  111. ### AI Config & Credential Files
  112. ```
  113. filetype:env "OPENAI_API_KEY" "KEYWORD"
  114. filetype:env "ANTHROPIC_API_KEY" "KEYWORD"
  115. filetype:env "HUGGINGFACE_TOKEN" "KEYWORD"
  116. filetype:env "GROQ_API_KEY" "KEYWORD"
  117. filetype:env "PINECONE_API_KEY" "KEYWORD"
  118. filetype:env "WANDB_API_KEY" "KEYWORD"
  119. filetype:env "DEEPSEEK_API_KEY" "KEYWORD"
  120. filetype:env "OPENROUTER_API_KEY" "KEYWORD"
  121. filetype:yaml "openai" "api_key" "KEYWORD"
  122. filetype:json "anthropic" "api_key" "KEYWORD"
  123. ```
  124. ### MCP Server Config Exposure (v1.2.0)
  125. ```
  126. # MCP server configs with embedded secrets (systemic RCE β€” Ox Security, April 2026)
  127. site:github.com "mcpServers" "args" filetype:json "KEYWORD"
  128. site:github.com "mcp.config" "apiKey" "KEYWORD"
  129. site:github.com ".cursor" "mcpServers" filetype:json "KEYWORD"
  130. site:github.com "windsurf" "mcp" "config" filetype:json "KEYWORD"
  131. # Claude Code leak artifacts
  132. site:github.com "claude-code" "leaked" "source" "KEYWORD"
  133. # VS Code YOLO mode (CVE-2025-53773)
  134. site:github.com "chat.tools.autoApprove" "true" filetype:json "KEYWORD"
  135. ```
  136. ---
  137. ## πŸ™ GitHub Dorks
  138. > Full collection: [`dorks/github/`](dorks/github/AI_GITHUB_DORKS.md)
  139. >
  140. > ⚠️ **Syntax note:** GitHub migrated to new Code Search. Use `path:*.env` instead of legacy `filename:.env`. Queries below use modern syntax where applicable.
  141. ### AI API Keys on GitHub
  142. ```
  143. # OpenAI (project keys β€” current format since April 2024)
  144. "sk-proj-" path:*.env "KEYWORD"
  145. "sk-proj-" path:*.py "KEYWORD"
  146. "OPENAI_API_KEY" path:*.env NOT "your_key" NOT "example" "KEYWORD"
  147. # Anthropic
  148. "sk-ant-api03" path:*.env "KEYWORD"
  149. "ANTHROPIC_API_KEY" path:*.env "KEYWORD"
  150. # Google AI / Gemini
  151. "AIzaSy" path:*.env "generativelanguage" "KEYWORD"
  152. "GOOGLE_API_KEY" path:*.env "gemini" "KEYWORD"
  153. # HuggingFace
  154. "hf_" path:*.env "KEYWORD"
  155. "HF_TOKEN" path:*.env
  156. # Groq
  157. "gsk_" path:*.env "KEYWORD"
  158. "GROQ_API_KEY" path:*.env "KEYWORD"
  159. # Replicate
  160. "r8_" path:*.env "KEYWORD"
  161. "REPLICATE_API_TOKEN" path:*.env "KEYWORD"
  162. # Vector DBs & MLOps
  163. "PINECONE_API_KEY" path:*.env "KEYWORD"
  164. "QDRANT_API_KEY" path:*.env "KEYWORD"
  165. "WANDB_API_KEY" path:*.env "KEYWORD"
  166. ```
  167. ### MCP & Agent Config Leaks
  168. ```
  169. path:mcp.json "api_key" "KEYWORD"
  170. path:mcp.json "KEYWORD"
  171. path:.cursor/mcp.json "KEYWORD"
  172. "mcpServers" path:*.json "apiKey" "KEYWORD"
  173. "mcpServers" path:*.json "OPENAI_API_KEY" "KEYWORD"
  174. ```
  175. ### System Prompts & Training Data
  176. ```
  177. "system_prompt" path:*.py "you are" "KEYWORD"
  178. "SYSTEM_PROMPT" path:*.env "KEYWORD"
  179. path:prompts.yaml "system" "KEYWORD"
  180. path:train.jsonl "prompt" "completion" "KEYWORD"
  181. path:dataset.jsonl "instruction" "output" "KEYWORD"
  182. ```
  183. ### MCP & AI IDE Config Exploitation (v1.2.0)
  184. ```
  185. # MCP STDIO configs (systemic RCE β€” 10+ CVEs)
  186. path:*.json "mcpServers" "command" "args" "KEYWORD"
  187. path:.vscode/settings.json "chat.tools.autoApprove" "true" "KEYWORD"
  188. # Claude Code attack vectors
  189. "CLAUDE.md" "permission" "allow"
  190. # DeepSeek keys
  191. path:*.env "DEEPSEEK_API_KEY" "KEYWORD"
  192. "DEEPSEEK_API_KEY" NOT "your_key" NOT "example"
  193. ```
  194. ---
  195. ## πŸ”­ Shodan Queries
  196. > Full collection: [`dorks/shodan/`](dorks/shodan/AI_SHODAN_DORKS.md)
  197. ### Self-Hosted LLMs
  198. ```
  199. # Ollama β€” 240,000+ exposed instances worldwide
  200. port:11434 product:"Ollama" "KEYWORD"
  201. port:11434 http.html:"Ollama" "KEYWORD"
  202. port:11434 "api/tags" "KEYWORD"
  203. # vLLM / OpenAI-compatible
  204. port:8000 "openai" "model" "KEYWORD"
  205. http.title:"FastAPI" port:8000 "/v1/models" "KEYWORD"
  206. # LM Studio
  207. port:1234 "/v1/models" "KEYWORD"
  208. # llama.cpp
  209. port:8080 "llama" "completion" "KEYWORD"
  210. ```
  211. ### AI Agent Gateways β€” CRITICAL
  212. ```
  213. # OpenClaw/Clawdbot β€” 4,000+ on Shodan, many with zero auth
  214. # Enables RCE via prompt injection, API key theft, reverse shells
  215. http.title:"Clawdbot Control" port:18789 "KEYWORD"
  216. http.title:"OpenClaw" port:18789 "KEYWORD"
  217. port:18789 "api/v1/status" "KEYWORD"
  218. port:18789 "auth_mode" "KEYWORD"
  219. ```
  220. ### Gradio & Streamlit Apps
  221. ```
  222. http.title:"Gradio" port:7860 "KEYWORD"
  223. http.title:"Streamlit" port:8501 "KEYWORD"
  224. http.title:"Stable Diffusion" port:7860 "KEYWORD"
  225. http.title:"ComfyUI" "KEYWORD"
  226. ```
  227. ### Vector Databases
  228. ```
  229. # Qdrant β€” NO auth by default
  230. port:6333 "qdrant" "KEYWORD"
  231. port:6333 "/collections" "KEYWORD"
  232. # Weaviate
  233. port:8080 "weaviate" "KEYWORD"
  234. # Milvus
  235. port:19530 "milvus" "KEYWORD"
  236. ```
  237. ### MLOps & Notebooks
  238. ```
  239. # MLflow β€” CVE-2026-0545 (CVSS 9.1) RCE, no auth by default
  240. http.title:"MLflow" port:5000 "KEYWORD"
  241. # Jupyter β€” ~10,000+ on Shodan, targeted by botnets
  242. http.title:"Jupyter Notebook" port:8888 -"Login" "KEYWORD"
  243. http.title:"JupyterLab" port:8888 "KEYWORD"
  244. # TensorBoard
  245. http.title:"TensorBoard" port:6006 "KEYWORD"
  246. ```
  247. ### MCP & AI Agent Gateways (v1.2.0)
  248. ```
  249. # MCP endpoints (systemic RCE β€” Ox Security)
  250. http.html:"mcp" "tools" port:3000 "KEYWORD"
  251. http.html:"Model Context Protocol" port:8080 "KEYWORD"
  252. # nginx-ui MCP (CVE-2026-33032 β€” actively exploited, CVSS 9.8)
  253. http.title:"Nginx UI" port:443 "KEYWORD"
  254. # Flowise (CVE-2026-40933)
  255. http.title:"Flowise" port:3000 "KEYWORD"
  256. # vLLM (LLMjacking target)
  257. http.html:"vLLM" port:8000 "KEYWORD"
  258. # Open WebUI
  259. http.title:"Open WebUI" port:3000 "KEYWORD"
  260. # LiteLLM Proxy
  261. http.title:"LiteLLM" port:4000 "KEYWORD"
  262. # DeepSeek-style ClickHouse exposure
  263. product:"ClickHouse" port:8123
  264. ```
  265. ---
  266. ## 🌐 Censys Queries
  267. > Full collection: [`dorks/censys/`](dorks/censys/AI_CENSYS_QUERIES.md)
  268. ```
  269. # Ollama (Censys found 25%+ on non-default ports)
  270. services.port=11434 AND services.http.response.body:"Ollama" "KEYWORD"
  271. # Gradio
  272. services.port=7860 AND services.http.response.html_title:"Gradio" "KEYWORD"
  273. # OpenClaw
  274. services.port=18789 AND services.http.response.body:"Clawdbot" "KEYWORD"
  275. # Vector DBs
  276. services.port=6333 AND services.http.response.body:"qdrant" "KEYWORD"
  277. services.port=8080 AND services.http.response.body:"weaviate" "KEYWORD"
  278. # MLOps
  279. services.port=5000 AND services.http.response.html_title:"MLflow" "KEYWORD"
  280. services.port=8888 AND services.http.response.html_title:"Jupyter" "KEYWORD"
  281. ### MCP & New AI Infrastructure (v1.2.0)
  282. # MCP endpoints
  283. services.http.response.body:"Model Context Protocol" AND services.port=3000 "KEYWORD"
  284. # Flowise (CVE-2026-40933)
  285. services.http.response.html_title:"Flowise" AND services.port=3000 "KEYWORD"
  286. # vLLM (LLMjacking target)
  287. services.http.response.body:"vLLM" AND services.port=8000 "KEYWORD"
  288. # nginx-ui MCP (CVE-2026-33032 β€” actively exploited)
  289. services.http.response.html_title:"Nginx UI" AND services.port=443 "KEYWORD"
  290. # ClickHouse exposure (DeepSeek-style)
  291. services.port=8123 AND services.http.response.body:"ClickHouse" "KEYWORD"
  292. # Open WebUI
  293. services.http.response.html_title:"Open WebUI" "KEYWORD"
  294. # LiteLLM
  295. services.http.response.html_title:"LiteLLM" "KEYWORD"
  296. ```
  297. ---
  298. ## 🌍 AI Service Endpoints
  299. ### Major Provider APIs
  300. | Provider | API Endpoint | Shared Content |
  301. |---|---|---|
  302. | **OpenAI** | `api.openai.com/v1/*` | `chatgpt.com/share/*` |
  303. | **Anthropic** | `api.anthropic.com/v1/*` | β€” |
  304. | **Google** | `generativelanguage.googleapis.com/v1beta/*` | β€” |
  305. | **xAI** | `api.x.ai/v1/*` | `grok.com/share/*` |
  306. | **Mistral** | `api.mistral.ai/v1/*` | β€” |
  307. | **Cohere** | `api.cohere.ai/v1/*` | β€” |
  308. | **DeepSeek** | `api.deepseek.com/v1/*` | β€” |
  309. | **Groq** | `api.groq.com/openai/v1/*` | β€” |
  310. | **HuggingFace** | `api-inference.huggingface.co/models/*` | `huggingface.co/spaces/*` |
  311. | **ElevenLabs** | `api.elevenlabs.io/v1/*` | β€” |
  312. ### Default Ports (all verified against official docs)
  313. | Service | Port | Auth Default | Risk Level |
  314. |---|---|---|---|
  315. | **Ollama** | 11434 | ❌ None | πŸ”΄ Critical |
  316. | **vLLM** | 8000 | ❌ None | πŸ”΄ Critical |
  317. | **OpenClaw** | 18789 | ❌ None (fixed in rebrand) | πŸ”΄ Critical |
  318. | **Gradio** | 7860 | ❌ None | πŸ”΄ Critical |
  319. | **Streamlit** | 8501 | ❌ None | 🟑 High |
  320. | **MLflow** | 5000 | ❌ None | πŸ”΄ Critical |
  321. | **Qdrant** | 6333/6334 | ❌ None | πŸ”΄ Critical |
  322. | **Weaviate** | 8080 (+ gRPC 50051) | ❌ None | 🟑 High |
  323. | **ChromaDB** | 8000 | ❌ None (binds localhost) | 🟑 High |
  324. | **Milvus** | 19530 (+ mgmt 9091) | ❌ None | 🟑 High |
  325. | **Jupyter** | 8888 | βœ… Token (often disabled) | 🟑 High |
  326. | **LM Studio** | 1234 | ❌ None | 🟑 High |
  327. | **GPT4All** | 4891 | ❌ None | 🟑 High |
  328. | **Flowise** | 3000 | ❌ None | πŸ”΄ Critical |
  329. | **LiteLLM Proxy** | 4000 | ❌ None | πŸ”΄ Critical |
  330. | **Open WebUI** | 3000/8080 | βœ… Auth (default) | 🟑 High |
  331. | **nginx-ui** | 443/9000 | βœ… Auth (bypassable) | πŸ”΄ Critical |
  332. | **ClickHouse** | 8123/9000 | ❌ None | πŸ”΄ Critical |
  333. | **TGI (HuggingFace)** | 8080 | ❌ None | 🟑 High |
  334. ---
  335. ## πŸ”‘ API Key Patterns
  336. ### Prefix Reference (verified against official docs)
  337. | Provider | Prefix | Length | Notes |
  338. |---|---|---|---|
  339. | **OpenAI** | `sk-proj-` | ~80+ chars | Current format since April 2024 |
  340. | **Anthropic** | `sk-ant-api03-` | ~90+ chars | API keys |
  341. | **Anthropic OAuth** | `sk-ant-oat01-` | β€” | OAuth tokens (Files API, etc.) |
  342. | **Google AI** | `AIzaSy` | 39 chars | ⚠️ Same prefix for Maps AND Gemini |
  343. | **HuggingFace** | `hf_` | ~34 chars | Read/write access tokens |
  344. | **Replicate** | `r8_` | ~40 chars | β€” |
  345. | **Groq** | `gsk_` | ~52 chars | β€” |
  346. ### Regex for Detection
  347. ```regex
  348. # OpenAI project keys
  349. sk-proj-[A-Za-z0-9_-]{80,}
  350. # Anthropic
  351. sk-ant-api03-[A-Za-z0-9_-]{90,}
  352. # HuggingFace
  353. hf_[A-Za-z0-9]{34}
  354. # Replicate
  355. r8_[A-Za-z0-9]{37}
  356. # Groq
  357. gsk_[A-Za-z0-9]{52}
  358. # Google AI (⚠️ also matches Maps, YouTube, etc.)
  359. AIzaSy[A-Za-z0-9_-]{33}
  360. ```
  361. > ⚠️ **Google `AIzaSy` warning:** Google uses the same prefix for ALL Cloud APIs. A Maps API key embedded in public JavaScript can silently gain Gemini API access if the Generative Language API is enabled on the same project. This was documented as a significant issue in early 2026.
  362. ---
  363. ## πŸ—ƒοΈ Vector Database Reconnaissance
  364. When exposed, vector databases leak proprietary embeddings, sensitive documents, and internal knowledge bases used in RAG systems.
  365. ### Enumeration Endpoints (verified against official API docs)
  366. | Database | List Collections | Extract Data |
  367. |---|---|---|
  368. | **Qdrant** | `GET /collections` | `POST /collections/{name}/points/scroll` |
  369. | **Weaviate** | `GET /v1/schema` | `GET /v1/objects?class={name}` |
  370. | **ChromaDB** | `GET /api/v2/collections` ⚠️ | `POST /api/v2/collections/{id}/query` |
  371. | **Milvus** | gRPC `ListCollections` | gRPC `Search/Query` |
  372. > ⚠️ **ChromaDB API update:** v1.0.0+ migrated to `/api/v2`. The legacy `/api/v1/collections` now returns a deprecation error. Update your recon scripts accordingly.
  373. ---
  374. ## πŸ€– MCP & AI Agent Exposure
  375. The **Model Context Protocol (MCP)** is the most critical emerging attack surface in AI security (2025-2026).
  376. ### Why MCP is Dangerous
  377. MCP servers connect AI models to shell access, file systems, databases, and APIs. Misconfigurations expose:
  378. - **Shell/code execution** via prompt injection
  379. - **API keys in plaintext** in `.env` files readable by agents
  380. - **Tool poisoning** β€” hidden instructions in tool metadata
  381. - **Supply chain attacks** β€” compromised MCP packages (24,000+ secrets leaked in MCP configs in its first year)
  382. ### Discovery Queries
  383. ```
  384. # Shodan
  385. port:18789 "api/v1" "KEYWORD"
  386. http.title:"Clawdbot" OR http.title:"OpenClaw" "KEYWORD"
  387. # GitHub
  388. path:mcp.json "apiKey" "KEYWORD"
  389. path:.cursor/mcp.json "KEYWORD"
  390. "mcpServers" path:*.json "KEYWORD"
  391. ```
  392. ### Key Incidents
  393. | Incident | Date | Impact |
  394. |---|---|---|
  395. | OpenClaw Shodan Exposure | Jan 2026 | 4,000+ agent gateways, many with zero auth and RCE |
  396. | Operation Bizarre Bazaar | Dec 2025–Jan 2026 | 35,000 attacks on LLM/MCP endpoints; commercial resale |
  397. | Smithery Registry Breach | 2025 | Fly.io token β†’ control of 3,000+ MCP servers |
  398. | mcp-remote CVE-2025-6514 | 2025 | Command injection in 437K+ installs |
  399. | Cursor IDE MCP Trust Issue | 2025 | Persistent RCE via shared repo configs (CVSS 7.2-8.8) |
  400. | MCP Systemic RCE (Ox Security) | Apr 2026 | 150M+ downloads, 200K servers, 10+ CVEs, 9/11 marketplaces poisoned |
  401. | nginx-ui MCPwn (CVE-2026-33032) | Mar 2026 | CVSS 9.8, actively exploited, full Nginx takeover in 2 requests |
  402. | Atlassian MCPwnfluence | Apr 2026 | CVE-2026-27825/27826 β€” RCE chain from LAN, no auth required |
  403. | MCP TypeScript SDK Data Leak | Apr 2026 | CVE-2026-25536 β€” cross-client data leak in shared McpServer instances |
  404. | Windsurf Zero-Click RCE | Apr 2026 | CVE-2026-30615 β€” zero-click prompt injection β†’ local RCE via MCP |
  405. > πŸ“– Full timeline: [`threat-intel/THREAT_INTELLIGENCE.md`](threat-intel/THREAT_INTELLIGENCE.md)
  406. ---
  407. ## πŸ“Š Threat Intelligence
  408. > Full entries: [`threat-intel/`](threat-intel/THREAT_INTELLIGENCE.md)
  409. - **Operation Bizarre Bazaar** β€” First large-scale LLMjacking: 35K attacks, commercial marketplace selling stolen AI access
  410. - **OpenClaw/Clawdbot Shodan Crisis** β€” CVE-2026-24061, "Localhost Trust" bypass
  411. - **ChatGPT/Grok Conversation Indexing** β€” Google indexed thousands of conversations with credentials
  412. - **Claude Conversation Indexing** β€” ~600 conversations indexed by Google (Forbes Sep 2025); 143K+ across all LLMs on Archive.org
  413. - **"Claudy Day" Attack Chain** β€” Open redirect + prompt injection + Files API exfiltration in claude.ai (Oasis Security, Mar 2026)
  414. - **Claude Code Source Map Leak** β€” 512K lines of source code exposed via npm package v2.1.88 (Mar 2026)
  415. - **MCP Supply Chain Timeline** β€” 10+ major breaches in MCP ecosystem
  416. - **175K Ollama Servers Exposed** β€” SentinelOne/Censys study across 130 countries
  417. - **AI-Assisted ICS Targeting** β€” 60+ Iranian groups using LLMs for critical infrastructure recon (Feb 2026)
  418. - **Claude Mythos Preview / Project Glasswing** β€” Anthropic's unreleased model autonomously discovered thousands of 0-days in major OS/browsers. Limited to ~50 organizations. CVE-2026-4747 (OpenBSD 27-year RCE).
  419. - **Claude Code Source Leak** β€” 59.8 MB source map published to npm (Mar 31, 2026). Critical 50-subcommand bypass vulnerability. Fake repos distributing Vidar/GhostSocks malware.
  420. - **MCP "Mother of All Supply Chains"** β€” Ox Security found architectural RCE in MCP SDKs: 150M+ downloads, 10+ CVEs, 9/11 marketplaces poisoned. Anthropic declined protocol fix.
  421. - **ChatGPT DNS Exfiltration** β€” Check Point discovered silent data leakage via DNS side channel in code execution sandbox. Patched Feb 20, 2026.
  422. - **CVE-2025-53773: Copilot Wormable RCE** β€” Prompt injection enables YOLO mode, wormable across repositories. Patched Aug 2025.
  423. - **OpenAI Codex CLI Command Injection** β€” Branch injection β†’ lateral movement β†’ codebase access. Patched Feb 2026.
  424. - **300K+ ChatGPT Credentials on Dark Web** β€” IBM X-Force 2026: AI chatbot credentials are emerging infostealer target.
  425. - **Ollama: 12,269 More Exposed** β€” LeakIX found additional exposed instances; maintainers continue rejecting auth PRs.
  426. - **nginx-ui MCPwn** β€” CVE-2026-33032 actively exploited. Full Nginx takeover in 2 HTTP requests.
  427. - **DeepSeek ClickHouse Exposure** β€” Wiz found 1M+ log entries with plaintext chats, API keys, backend details (Jan 2025).
  428. ---
  429. ## πŸ› οΈ Tools
  430. > Only AI-specific tools. Full details: [`tools/TOOLS.md`](tools/TOOLS.md)
  431. > For generic OSINT tools (Shodan, Censys, etc.) see [cloud_osint](https://github.com/7WaySecurity/cloud_osint).
  432. | Tool | What It Does | By |
  433. |---|---|---|
  434. | [**Garak**](https://github.com/NVIDIA/garak) | LLM vulnerability scanner β€” "nmap for LLMs" | NVIDIA |
  435. | [**PyRIT**](https://github.com/microsoft/PyRIT) | AI red teaming framework β€” Crescendo, jailbreaking, multi-turn | Microsoft |
  436. | [**promptfoo**](https://github.com/promptfoo/promptfoo) | LLM pentesting CLI β€” 133+ attack plugins | OpenAI |
  437. | [**DeepTeam**](https://github.com/confident-ai/deepteam) | 50+ vulns, 20+ attacks, OWASP/MITRE mapping | Confident AI |
  438. | [**API Radar**](https://apiradar.live/) | Real-time leaked AI API key monitoring on GitHub | Independent |
  439. | [**KeyLeak Detector**](https://github.com/Amal-David/keyleak-detector) | Web scanner for 200+ patterns incl. 15+ AI providers | Independent |
  440. | [**promptmap**](https://github.com/utkusen/promptmap) | ChatGPT dorks & prompt injection testing | Utku Şen |
  441. | [**Vulnerable MCP**](https://vulnerablemcp.info/) | MCP vulnerability database with CVEs and PoCs | Community |
  442. | [**MCPSafetyScanner**](https://github.com/johnhalloran321/mcpSafetyScanner) | MCP server security auditing tool | Academic (Leidos) |
  443. | [**Cisco AI Supply Chain Scanners**](https://blogs.cisco.com/ai/cisco-state-of-ai-security-2026-report) | Scanners for MCP, A2A, pickle files, agentic skill files | Cisco Talos |
  444. | [**DorkEye**](https://github.com/search?q=DorkEye+OSINT) | Automated Google Dorking with multi-agent analysis pipeline | Open Source |
  445. ---
  446. ## 🚨 Detection Rules
  447. > Full Sigma rules: [`detection-rules/SIGMA_RULES.md`](detection-rules/SIGMA_RULES.md)
  448. 12 Sigma detection rules covering:
  449. - External access to Ollama (port 11434)
  450. - AI API keys appearing in application logs
  451. - Unauthorized LLM API access without Bearer tokens
  452. - Exposed MCP servers with `auth_mode: none`
  453. - Vector database unauthorized enumeration
  454. - LLMjacking β€” anomalous inference spikes (>500 requests/hour)
  455. - AI agent RCE via prompt injection (suspicious tool calls to bash/python_repl)
  456. - MCP STDIO arbitrary command execution (CVE-2026-30615/30624/30616/40933)
  457. - VS Code Copilot YOLO mode activation (CVE-2025-53773)
  458. - nginx-ui MCP endpoint unauthenticated access (CVE-2026-33032, actively exploited)
  459. - DNS-based data exfiltration from AI code execution sandboxes
  460. - Claude Code 50+ subcommand pipeline deny rule bypass
  461. ---
  462. ## πŸ“š References
  463. ### Research
  464. | Source | Finding | Link |
  465. |---|---|---|
  466. | SentinelOne + Censys | 175,108 exposed Ollama hosts across 130 countries | [SecurityWeek](https://www.securityweek.com/175000-exposed-ollama-hosts-could-enable-llm-abuse/) |
  467. | Cisco Talos | 1,100+ Ollama servers, 20% serving models without auth | [Cisco Blog](https://blogs.cisco.com/security/detecting-exposed-llm-servers-shodan-case-study-on-ollama) |
  468. | Pillar Security | Operation Bizarre Bazaar: 35,000 LLMjacking attacks | [Pillar Security](https://www.pillar.security/) |
  469. | GitGuardian | AI credential leaks surged 81% YoY; 29M secrets on GitHub | [Report](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/) |
  470. | AuthZed | Complete MCP security breaches timeline | [Blog](https://authzed.com/blog/timeline-mcp-breaches) |
  471. | Pangea | Sensitive data in indexed ChatGPT histories | [Blog](https://pangea.cloud/blog/mining-the-index-uncovering-sensitive-data-in-public-chatgpt-histories-via-google-search/) |
  472. | Trail of Bits | 8 high-severity vulns in Gradio 5 audit | [Blog](https://blog.trailofbits.com/2024/10/10/auditing-gradio-5-hugging-faces-ml-gui-framework/) |
  473. | Oasis Security | "Claudy Day" β€” 3 vulns chained for data exfiltration from Claude | [Blog](https://www.oasis.security/blog/claude-ai-prompt-injection-data-exfiltration-vulnerability) |
  474. | Forbes | ~600 Claude conversations indexed by Google | [Article](https://www.forbes.com/sites/iainmartin/2025/09/08/hundreds-of-anthropic-chatbot-transcripts-showed-up-in-google-search/) |
  475. | Obsidian Security | 143K+ LLM chats (incl. Claude) on Archive.org | [Blog](https://www.obsidiansecurity.com/resource/143k-claude-copilot-chatgpt-chats-publicly-accessible-were-you-exposed) |
  476. | TechRadar | Claude Code 512K source lines leaked via npm | [Article](https://www.techradar.com/pro/security/anthropic-confirms-it-leaked-512-000-lines-of-claude-code-source-code-spilling-some-of-its-biggest-secrets) |
  477. | Penligent | Claude Code source map leak analysis | [Blog](https://www.penligent.ai/hackinglabs/claude-code-source-map-leak-what-was-exposed-and-what-it-means/) |
  478. | Ox Security | 150M+ MCP downloads affected by systemic RCE, 10+ CVEs, 9/11 marketplaces poisoned | [Blog](https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-critical-systemic-vulnerability-at-the-core-of-the-mcp/) |
  479. | Anthropic Red Team | Claude Mythos: thousands of 0-days, CVE-2026-4747, sandbox escape | [red.anthropic.com](https://red.anthropic.com/2026/mythos-preview/) |
  480. | IBM X-Force | 300K+ ChatGPT creds on dark web, supply chain attacks 4x in 5 years | [Report](https://www.ibm.com/think/insights/more-2026-cyberthreat-trends) |
  481. | Check Point | ChatGPT DNS exfiltration β€” silent data leakage via side channel | [Blog](https://blog.checkpoint.com/research/when-ai-trust-breaks-the-chatgpt-data-leakage-flaw-that-redefined-ai-vendor-security-trust/) |
  482. | Zscaler ThreatLabz | Claude Code source leak + Vidar/GhostSocks malware lure | [Blog](https://www.zscaler.com/blogs/security-research/anthropic-claude-code-leak) |
  483. | Cisco AI | State of AI Security 2026 β€” MCP, A2A, agentic AI scanners | [Report](https://blogs.cisco.com/ai/cisco-state-of-ai-security-2026-report) |
  484. | LeakIX | 12,269 additional exposed Ollama instances, auth PR rejection critique | [Blog](https://blog.leakix.net/2026/02/ollama-exposed/) |
  485. | CSA | Time-to-exploit now under 20 hours (Mythos briefing) | [HelpNetSecurity](https://www.helpnetsecurity.com/2026/04/15/anthropic-claude-mythos-ai-vulnerability-discovery/) |
  486. | Wiz | Claude Mythos practical guidance β€” "Y2K moment" for cybersecurity | [Blog](https://www.wiz.io/blog/claude-mythos) |
  487. | Wiz | DeepSeek ClickHouse exposure β€” 1M+ log entries with plaintext chats | [Blog](https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak) |
  488. ### Standards
  489. - [OWASP Top 10 for LLM Applications 2025](https://owasp.org/www-project-top-10-for-large-language-model-applications/)
  490. - [OWASP Top 10 for Agentic Applications 2026](https://owasp.org/www-project-top-10-for-agentic-applications/)
  491. - [MITRE ATLAS](https://atlas.mitre.org/)
  492. - [MCP Security Best Practices](https://modelcontextprotocol.io/specification/draft/basic/security_best_practices)
  493. - [vulnerablemcp.info](https://vulnerablemcp.info/)
  494. ---
  495. ## ⚠️ Disclaimer
  496. The resources in this repository are for **authorized security testing, education, and legitimate research only**.
  497. 1. **Obtain proper authorization** before testing infrastructure you do not own
  498. 2. **Follow applicable laws** (CFAA, GDPR, local equivalents)
  499. 3. **Report vulnerabilities responsibly** through appropriate disclosure channels
  500. 4. **Never exploit** discovered misconfigurations for unauthorized access
  501. The authors assume no liability for misuse.
  502. ---
  503. ## 🀝 Contributing
  504. PRs welcome! Please verify dorks/queries work before submitting. See [CONTRIBUTING.md](CONTRIBUTING.md).
  505. ---
  506. ## License
  507. Code: [MIT](LICENSE) Β· Data & Documentation: [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)
  508. ---
  509. <p align="center"><b>Maintained by <a href="https://7waysecurity.co">7WaySecurity</a></b> Β· Last updated June 2026 Β· v1.4.0</p>

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.